JP2011141666A - Device, method and program for recommending interest information - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an interest information recommendation device clearly indicating a cluster from which a content recommended to an active user a is found. <P>SOLUTION: The interest information recommendation device includes: a user interest similarity measurement instrument 10 measuring interest similarity of each user based on taxonomy in consideration of a class including a word inside interest information of each user, an interest degree of each instance, and topology that is a connection form of the class; a user interest correlation graph generator 20 setting each user as a node based on the user interest similarity, setting an edge having a weight based on the similarity as a score between the users, and generating a user interest correlation graph; a recommendation score calculator 30 randomly walking on the graph, and calculating a recommendation score of an interest target content; and a belonging cluster specification device 40 of a recommendation item clustering the graph, and creating a recommendation list including the recommendation item and background information of the item based on a result thereof. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to content recommendation based on collaborative filtering, and in particular to graph-based collaborative filtering, and more particularly to a user interest matching algorithm based on taxonomies to improve recommendation accuracy for users with various interests.

  Non-patent document 1 is known as a conventional graph-based collaborative filtering technique. This is mainly due to the problem of sparseness problem (sparse problem) that the user can not make a good recommendation when content consumption (use) is small. By pursuing the content of interest of the user who is a little far away and using it for recommendation, the parity problem is improved.

  As a similar approach, a technique using a random walk (see Non-Patent Document 2) has been frequently used in recent years. In that method, an item graph based on item similarity (items are nodes, edges are between similar items. Edges are given a score based on similarity), and the interest of the active user a is first created. Based on the number of appearances of the item of interest that the user has followed and the edge score, the item's recommended value is calculated based on the item as the starting point, and the item's recommended value is ranked based on the recommended value. indicate.

  It is known that this method can also handle the parity problem (according to this non-patent document 2, the method is applied to the item graph because the item generally uses the information recommendation system. It is pointed out that it is less than the number of (but the current system has many opposite cases).

  However, the above-described method has the following two drawbacks.

  (1) Conventionally, in the construction of the user graph based on the similarity of the user's interest, the similarity calculation between users using only the user's consumed content is performed. In that case, the same content is consumed (used), and even if the interest is close to some extent, the similarity often becomes 0 unless both the same content is consumed. Therefore, even if the interest is close to some extent, the edge is not drawn (or the edge score is 0) in many cases. Therefore, when a random walk is performed on the item graph, there are many users who do not follow immediately even though their interest is close to some extent, and the accuracy of ranking is lowered.

  In addition, as a result of the same cause, consuming the same content means that the users are quite similar, so only those content with the same tendency can be found from such users, and a wide range of interests can be found. The accuracy of the recommendation to users with is reduced.

  In summary, if the information recommendation system knows only a part of the user's various interests, and the remaining part is different from the user's interest that the system knows, the recommendation will not be successful. . This problem is very similar to the parity problem in that a user having an item that the user may be interested in does not appear around the user.

  (2) Conventionally, the user graph is randomly walked in the vertical and horizontal directions, and it is not possible to grasp the relationship between the followed user and the active user a. For this reason, it is difficult to grasp what kind of content has been found from which user, so it is difficult to grasp the characteristics of the found content.

  Also, in Patent Document 1, as a WEB page that reflects the user's interest, blog entries such as NTT Resonant's goo blog and NTT data doblog are different from conventional bulletin boards, and a specific individual person feels his own. This information is written and spelled, so this information contains a lot of information that expresses the interests of individuals, and this blog entry has a taxonomy class (genre etc.) provided by each service provider. ), A technique that enables the user's interest to be extracted along the taxonomy is disclosed.

  However, in Patent Document 1, the user interest along the taxonomy is not necessarily limited to that extracted from the blog, but can be extracted from the purchase history and the music audition history.

  Patent Document 2 discloses a method of giving an interest level, which is a strength of a user's interest in an object of interest (instance, class) when creating the user's interest along a taxonomy.

  With such a technical background, a method of extracting user interest along a taxonomy is being given.

  Further, Patent Literature 3 discloses a method of calculating the similarity of interest between users from the similarity of user interest along the taxonomy.

  In addition, the thing of a nonpatent literature 3-5 is well-known as a technique relevant to this invention.

JP 2007-226468 A JP 2008-27142 A JP 2008-27143 A

ZAN HUANG, HSINCHUN, and DANIEL ZENG, “Applying associate retrievable technologies to alleviate the inspirables in the CV”. 22, N0.1, January 2004, pp. 116-142 Hilmi Yildirim and Mukakai S. Krishnamorthy, “A Random Walk Method for Allivating the Sparity Problem in Collaborative Filtering”, SIG-REC2008 Fast algorithm for detecting community structure in networks Export by: M.M. E. J. et al. Newman, Internet <URL: http: // arxiv. org / abs / cond-mat / 3095098>, [retrieval on December 22, 2009] Makoto Nakajo, Yu Miyoshi, Yoshihiro Otsuka, “Proposal of Automatic Generation of User Interest Ontology Based on Blog Data and Supporting Community Formation”, DEWS 2006 2A-i7, 2006 Toshiyuki Masui, “City Corner of Interface (93)-Bookshelf Calculation”, UNIX MAGAZINE 2005.12

  In the present invention, based on the technical background, a user interest graph is created and a random walk is performed on the graph in order to deal with the problems (1) and (2) described above. The object is to provide an apparatus, a method, and a program that clearly indicate from which cluster the content recommended for a is found.

The features of the present invention are the following two points.
(1) Contents that complement user interest extraction (Patent Document 1) and similarity calculation (Patent Document 3) along the taxonomy proposed by the inventor in the past. The above-described method is configured to cover an area in which the user interest similarity graph along the taxonomy can be generated, the random walk on the graph can be applied, and the sparse problem can be improved thereby.
(2) A new method for clustering user graphs was constructed. Furthermore, a method has been constructed to clearly indicate what kind of cluster the content recommended for the active user a comes from when considering the clustering result of the graph.

  In other words, in the present invention, the interest similarity between users is calculated using not only the content but also the taxonomy that is the classification knowledge of the content, and the user graph is constructed. Increase accuracy. By using a taxonomy, similarity can be determined not only for users who are interested in the same instance but also among users who are interested in the same class, so it is particularly effective for the Sparsity problem (problem (1)) Correspondence). Furthermore, using a prior result obtained by clustering the user graph, it is possible to recommend content that is highly likely to be of interest to the user although it cannot be found from the cluster to which the user graph belongs (corresponding to the problem (2)).

  An interest information recommendation device according to the present invention generates a personal ontology that conceptually hires interest information constructed for each user, and includes a class including words in the interest information, an interest degree for each instance, and a connection form of classes. Considering a certain topology, based on the user interest similarity measurer that measures the interest similarity for each user based on the taxonomy, and based on the user interest similarity measure for each user measured by the user interest similarity measurer User interest correlation graph generation means for generating a user interest correlation graph by setting an edge having a user as a node and a weight based on similarity as a score between the users, and a graph generated by the user interest correlation graph generation means Randomly walk up and add the content of interest of the user who followed the content to the recommendation target list. Clustering the recommendation score calculation means for calculating the recommendation score of the content of interest based on the count number of the user and the score of the edge between users, and the graph generated by the user interest correlation graph generation means. And a recommended item affiliation cluster specifying means for creating a recommended list including a recommended item and background information of the item.

(1) According to the inventions described in claims 1 to 9, the recommendation accuracy can be improved particularly in a stage where the user's interest is not sufficiently grasped by the information recommendation system, for example.

  Even if a recommended item does not exist in a cluster to which a user belongs (ie, an item that is not of interest to the user so far), for example, information on which cluster it was found from (recommended item) The user can clearly grasp what kind of content has been discovered by which user, and can easily grasp the characteristics of the content.

For this reason, unlike the collaborative filtering performed in conventional Amazon, etc., unlike the recommendation list in which similar instances are arranged, the instances coming out of various separate clusters are separated into the recommendation list while being separated for each cluster. By mixing, it is possible to create a recommendation list of a variety of instances that the user has a high probability of being interested in and present to the user.
(2) According to the inventions of claims 2 and 6, the interest coincidence S (T ₁ ) for the common class set forming the topology T ₁ composed of the parent class and the child class along the taxonomy, the class and the instance The degree of approximation So between the ontology is obtained by the interest coincidence S (T ₂ ) for the common class set forming the topology T ₂ consisting of and the evaluation function f (X) according to the importance for the topologies T ₁ and T ₂ . Even if the information recommendation system does not fully grasp the user's interest, for example, if there is a user who connects not only with the instance but also with the class, the edge may be drawn, and the recommendation error is reduced. Can be suppressed.
(3) According to the inventions according to claims 3 and 7, since the number of users X for setting an edge is determined, a user interest correlation graph capable of correctly obtaining a user set having similar interests Can be generated.
(4) According to the inventions described in claims 4 and 8, unlike the past, the recommended information is not from the cluster to which the user belongs. It becomes possible to make a recommendation while clearly stating that the instance is out of ambiguity.

  Also, by specifying the discovery base cluster of recommended instances, for example, when instances have been recommended from a plurality of clusters, the clusters for each recommended instance can be specified. If the user has jumped, the user can be recommended while adding information that the sources are different.

The block diagram which shows the example of embodiment of the interest information recommendation apparatus of this invention. The flowchart which shows the embodiment of the interest information recommendation method of this invention. Explanatory drawing which shows the procedure of the user distribution analysis and interest ontology automatic generation which are utilized in the embodiment of this invention. Explanatory drawing of the approximation measurement algorithm in patent document 3 utilized in the embodiment of this invention. Explanatory drawing which shows the example of the degree of interest in the embodiment example of this invention. Explanatory drawing of the approximation degree measurement algorithm between ontology in consideration of the degree of interest in the embodiment of this invention. Explanatory drawing of the realization service image by the approximation degree measurement of interest ontology in the example embodiment of this invention. The flowchart which shows the detail of a process of the user interest correlation graph production | generation step of FIG. The flowchart which shows the detail of a process of the affiliation cluster identification step of the recommendation item of FIG. Explanatory drawing which shows an example of the recommendation list | wrist output by this invention. The graph which shows the comparison verification of this invention and a prior art, and shows the recommendation error with respect to the user interest grasp degree of a recommendation system.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings, but the present invention is not limited to the following embodiments.

  FIG. 1 is a block diagram showing a schematic configuration of a system to which an interest information recommendation device according to an embodiment of the present invention is applied. Reference numeral 1 denotes a communication network such as the Internet or an IP-VPN (Internet Protocol-Virtual Private Network), and the terminals 2 to 4 and the server 5 are connected via the communication network 1.

  The terminals 2 to 4 are notebook personal computers or desktop personal computers, and may be mobile phone terminals or PDAs (Personal Data Assistants). The server 5 can be installed on a blog provider or ISP (Internet Service Provider).

  Here, the server 5 is provided with blog sites 7 to 9 corresponding to the terminals 2 to 4 respectively, and the blog entries 7a to 7n, 8a to 8n, and 9a to 9n are held in the blog sites 7 to 9, respectively. Has been. The blog entries 7a to 7n, 8a to 8n, and 9a to 9n represent minimum units of articles in the blog, and can be provided for each day. The server 5 holds a template ontology 6, and the template ontology 6 is provided with a personal ontology template in which personal (user) interest information is conceptually hierarchized.

  The template ontology 6 can be arbitrarily created on the blog provider side. For example, if the blog provider wants the users of the terminals 2 to 4 to construct a personal ontology related to music, the template ontology 6 related to music may be constructed. Here, in order to express the interests of the users of the terminals 2 to 4 in detail, it is preferable to construct a template ontology 6 that is subdivided as much as possible and has high coverage. The entity of the template ontology 6 is a representation of a class / instance tree on the DB, such as a text file described in an XML language such as the ontology description language OWL. In order to simplify the organization of information, instances may be classified only into the lowest class.

  Further, the server 5 includes a user interest similarity measurer 10 as a user interest similarity measurer, a user interest correlation graph generator 20 as a user interest correlation graph generator, a recommended score calculator 30 as a recommended score calculator, and A recommended item belonging cluster identification device 40 is provided as a recommended item belonging cluster identifying means.

  Each process executed by the user interest similarity measuring device 10, the user interest correlation graph generator 20, the recommendation score calculator 30, and the recommended item affiliation cluster specifying device 40 is performed according to the flowchart of FIG. Each function is achieved by a computer, for example.

<User Interest Similarity Measuring Device 10 (Step S100)>
(1) The user interest similarity measurer 10 receives user interest, that is, information on the relationship between the user and the item such as purchase history and viewing history, and analyzes the interest distribution analysis and interest of the user described in Patent Documents 1 and 3, for example. Ontology automatic generation (FIGS. 3 to 5) and interest similarity measurement for each user (FIGS. 4 to 7) are performed as follows.

  (1-1) First, in FIG. 3 showing a user distribution analysis and an interest ontology automatic generation procedure, all blog entries (users) collected through a ping server (server 5 in FIG. 1) that collects and provides blog update information, etc. Morphological analysis is performed on the entry set of A to X), and words that frequently appear in the entry are extracted to create an index (step S1). Here, it is assumed that the collected blog entries have a unique user ID.

(1-2) Next, all blog entries are classified with respect to the template ontology OH (step S2). As the classification method, if there is a name attribute of the class C _i with stationery ontology OH to the description in a certain entry, classifies the entry to C _i, also, instance I _i belonging to C _i (∈C _i) If there is a name attribute, classify the entry into an instance I _i of class C _i . Note that the same entry may be classified into a plurality of classes. For example, in FIG. 3, when there is a character string “Charlantans” in the description in the entry, the entry is classified into an instance “Charlantans” of the class “Machchester”.

(1-3) Next, the number of users who are interested in each instance of the lowermost class C ₁ forming the template ontology OH is measured (step S3). When calculating the number of users interested in the instance I _l of the class C _l, the number of users is measured as 1 even if the same user describes the instance I _{l in a} plurality of entries. Then also performed on the lowermost class the measurement, the number of users who are interested in the lowest layer class, interested in the number of users and the lowermost class C _l itself interested in all instances of subordinate lowermost class Measure with the total number of users. Also in this case, even if the same user is interested in a plurality of instances or is interested in the lowermost class and instances belonging to the class at the same time, the number of users is measured as one. In this way, by recursively measuring the number of users interested in the classes and instances forming the template ontology OH up to the root class, the distribution of users interested in the domain can be measured.

  (1-4) Next, after classifying all entries, the classification result is arranged for each user ID, and an interest ontology for each user A to X is generated.

  (2) Next, the similarity measurement algorithm described in Patent Document 3 will be described with reference to FIG.

  In the following description, when the degree of approximation between an ontology and another ontology is measured, the former is called a source ontology and the latter is called a target ontology.

  In FIG. 4, classes “b1”, “b2”, and “b3” exist immediately below the class “a1” of the template ontology OH, and “c1” and “c3” immediately below the class “b1”. It is assumed that a class called “c2” exists and classes “d1” and “d2” exist immediately below the class called “c1”.

  Also, it is assumed that classes “c3” and “c4” exist immediately below the class “b2”, and a class “c5” exists immediately below the class “b3”.

  The class “d1” includes instances “j” and “k”, the class “d2” includes the instance “l”, and the class “c2” includes “m”. "B", the class "b2" has an instance "n", and the class "c3" has "a", "e", "c", "f", "b" ”,“ D ”, and“ g ”exist, and the class“ c4 ”includes instances“ p ”,“ q ”,“ j ”, and“ h ”.

  Then, the words that frequently appear in each user's blog entry are extracted, and the classes or instances including the words and all the classes above them are extracted from the template ontology OH, thereby creating personal ontologies OA and OB. Shall be.

  Here, the classes “b1” and “b2” exist immediately below the class “a1” of the personal ontology OA, and the classes “c1” and “c2” exist immediately below the class “b1”. It is assumed that a class “d1” exists immediately below the class “c1”, and a class “c3” and “c4” exist immediately below the class “b2”. , The class “d1” includes instances “j” and “k”, the class “c2” includes the instance “m”, and the class “c3” includes “a”. , “C”, “b”, and “d” exist, and the class “c4” includes instances “q” and “h”.

  Also, in the personal ontology OB, there are “b1”, “b2”, and “b3” classes directly under the “a1” class, and there is a “c1” class directly under the “b1” class. Assume that a class “d2” exists immediately under the class “c1”, and a class “c3” and “c4” exist immediately below the class “b2”. The class “d2” has an instance “l”, the class “b2” has an instance “n”, and the class “c3” has “a” and “c”. ”,“ E ”, and“ f ”exist, and“ c4 ”class includes“ p ”and“ j ”.

  (2-1) First, a common class excluding a terminal class is analyzed between the template ontology OH and the personal ontology OA and OB, and a topology composed of parent-child classes with the common class as a parent class is extracted. In the example of FIG. 4, the end classes are “d1”, “d2”, “c3”, and “c4”. As a result, a child class set G1 having the class “a1” as the parent class, a child class set G2 having the class “b1” as the parent class, and the class “c1” being the parent class between the template ontology OH and the personal ontologies OA and OB. Child class set G3, child class set G4 having class “b2” as a parent class, and child class set G5 having class “b3” as a parent class can be extracted.

  The analysis of the common class only needs to confirm whether the same class ID exists between the template ontology OH and the personal ontology OA, OB. For this reason, it is not necessary to measure the degree of approximation of class name attributes, instance set properties, etc., and the amount of calculation can be reduced while maintaining the correspondence between classes accurately.

(2-2) Next, the degree of approximation between the child class sets X and Y forming each topology between the personal ontology OA and OB is calculated with depth priority. Here, if the set of child classes of the classes constituting the template ontology OH is Z, the degree of approximation between the child class sets X and Y can be obtained by | X∩Y | / | Z |. By adding the approximation degree between the child class set for each topology, it is possible to measure the degree of approximation S _T personal ontology OA, between OB topology.

  For example, in the child class set G1, since the template ontology OH includes child classes “b1”, “b2”, and “b3”, the number of members of the child class set of the template ontology OH in the child class set G1 is 3 In the child class set G1, the personal ontology OA includes child classes “b1” and “b2”, and the personal ontology OB includes child classes “b1”, “b2”, and “b3”. Therefore, the child classes that are commonly included in the personal ontologies OA and OB are only “b1” and “b2”, and the number of members of the product set of the child class sets of the personal ontology OA and OB in the child class set G1 is 2. As a result, the degree of approximation between the child ontology sets G1 of the personal ontology OA and OB is 2/3.

Similarly, the degree of approximation between the personal ontology OA and OB between the child class sets G2 is 1/2, the degree of approximation between the child class sets G3 is 0/2, and the degree of approximation between the child class sets G4 is 2/2. As a result, personal ontology OA, similarity S _T topology between OB becomes 2/3 + 1/2 + 0/2 + 2/2.

  In the calculation of the degree of topology approximation between the personal ontology OA and OB, the class ID assigned from the template ontology OH is referred to, and the one along the connection form of the template ontology OH exists between the personal ontology OA and OB. You only have to confirm whether you want to do it. For example, whether the connection form a1-b1-c1 in the personal ontology OA, OB is the same as the connection form of the template ontology OH depends on whether the classes “a1”, “b1”, and “c1” are the personal ontology OA, OB. It can be determined simply by confirming whether or not it is held at.

  For this reason, in order to confirm the degree of coincidence of the topology between the personal ontology OA and OB, it is only necessary to check the class ID possessed by the personal ontology OA and OB. Therefore, it is possible to reduce the calculation amount of the degree of approximation of the topology between the personal ontology OA and OB.

(2-3) Next, the degree of approximation between the common classes of personal ontology OA and OB is calculated. Here, when calculating the degree of approximation between the common classes of the personal ontology OA and OB, an instance set belonging to the class can be used. That is, in a certain class C1, if the instance set of the source ontology is x, the instance set of the target ontology is y, and the instance set of the template ontology OH is z, the degree of approximation between the common classes of the personal ontology OA and OB is | x ∩y | / | z |. By adding the degree of approximation between the common class, it is possible to measure the degree of approximation S _C between personal ontology OA, the OB classes.

  For example, in the common class “b2” between the personal ontologies OA and OB, the instance set G6 in the common class “b2” of the personal ontology OA includes the instance “n”, and the common class “b2” of the personal ontology OB. An instance “n” exists in the instance set G6, and an instance “n” exists in the common class “b2” of the template ontology OH. As a result, the number of members of the common class “b2” of the template ontology OH is 1, the number of members of the product set of the instance set G6 of the personal ontologies OA and OB is 1, and the degree of approximation between the common classes “b2” is 1/1.

  Further, in the common class “c3” between the personal ontology OA and OB, the instance set G7 in the common class “c3” of the personal ontology OA includes instances “a”, “c”, “b”, and “d”. In the instance set G7 in the common class “c3” of the personal ontology OB, instances “a”, “c”, “e”, and “f” exist, and in the common class “c3” of the template ontology OH “A”, “e”, “c”, “f”, “b”, “d” and “g” exist.

As a result, the number of members of the common class “c3” of the template ontology OH is 7, the number of members of the product set of the instance set G7 of the personal ontologies OA and OB is 2, and the degree of approximation between the common classes “c3” is 2/7. Accordingly, the degree of approximation S _C between the classes of the personal ontology OA and OB is 1/1 + 2/7.

  Note that even in the calculation of the degree of approximation between the common classes of the personal ontology OA and OB, it is only necessary to confirm whether the instance ID allocated from the template ontology OH exists between the personal ontology OA and OB. For this reason, in order to calculate the degree of approximation between the classes of the personal ontology OA and OB, it is not necessary to confirm in advance the correspondence between instances based on the consistency of the names of instances, and the amount of calculation can be reduced.

(2-4) Then, when the topology approximation S _T and the class approximation S _C in the personal ontology OA, OB are obtained, the evaluation function f (X) corresponding to the importance for the topology and the class is used. The approximation S _O (AB) between the personal ontology OA and OB can be given by the following equation (1).

S _O (AB) = S _T + f (S _C ) (1)
Also, by introducing an interest level that indicates the degree of interest in the class / instance that constitutes the interest ontology for each user, the user's detailed interest is reflected, and even among users who are interested in the same class / instance, Try to create a community between nearby users. First, the definition of interest is given.

  (2-5-1) The degree of interest of the user per entry is 1.

(2-5-2) When the types of user classes and instances appearing in an entry E _i are N (E _i ), the degree of interest of each class and instance of the user in the entry E _i is 1 / N ( E _i ).

(2-5-3) The degree of interest for each instance I _i in the ontology is given by the following equation (2), where E is the total accumulated entry set of the user.

(2-5-4) The degree of interest for each class C _i in the ontology is given by the following equation (3), where E is the total accumulated entry set of the user.

  (2-5-5) The interest level of the instance is inherited by the interest level of the class to which the instance belongs, and the interest level of the child class is inherited by the interest level of the parent class.

  FIG. 5 is a diagram showing an example of the degree of interest of the classes and instances in FIGS. In FIG. 5, it is assumed that entries A1 and E2 are included in all the stored entries of user A. In the interest ontology OA of the user A, a class called “alternative” exists, and a class called “alternative” has child classes called “Madchester” and “Shoegaze”. The class “alternative” has an instance “nirvana”, the class “Madchester” has instances “New Order” and “Stone Roses”, and the class “Shoegaze” has “ Assume that instances of “My Bloody Valentine” and “Ride” exist.

  In entry E1, four artists “Stone Roses”, “My Bloody Valentine”, “nirvana”, and “New Order” appear, and in entry E2, one artist “My Bloody Valentine” appears again. , One genre of “Shoegaze” and one artist of “Ride” have appeared.

  Therefore, the interest level of the instance “New Order” is 1/4, the interest level of the instance “Stone Roses” is 1/4, and the interest level of the instance “My Bloody Valentine” is 1/4 + 1/3 = 7 / 12. The interest level of the instance “Ride” is 1/3, and the interest level of the instance “nirvana” is 1/4.

  The interest level of the class “Madchester” can be obtained by adding the interest levels of the instances “New Order” and “Stone Roses” under the class, and 1/4 + 1/4 = 1/2. It becomes. Further, the interest level of the class “Shoegaze” can be obtained by adding the interest levels of the instances “My Bloody Valentine” and “Ride” under the class, and 7/12 + 1/3 + 1/3 = 5 / 4. Furthermore, the interest level of the class “alternative” can be obtained by adding the interest levels of the instances “nirvana”, “Madchester”, and “Shoegaze” under the class, 1/2 + 5/4 + 1 / 4 = 2.

  As a result, the artist “My Bloody Valentine” has been written in the two entries E <b> 1 and E <b> 2, so that it can be recognized that the user A is highly interested. In addition, for the genre of “Shoegaze”, the class name also appears in entry E2, and the subordinate artists are frequently spoken in entries E1 and E2, so the degree of interest can be increased, and simply Even if there are many artists under them, it can be prevented that they are treated equally as interests.

  (3) Next, an approximation measurement algorithm, which is a method of the present invention for obtaining the approximation of ontology questions in consideration of the interest for each class / instance and the topology that is the class connection form, will be described with reference to FIG.

  In FIG. 6, the classes “b1” and “b2” exist immediately below the class “a1” of the interest ontology OA of the user A, and “c3” and “c4” exist immediately below the class “b2”. The class “b1” has instances “m” and “k”, and the class “b2” has an instance “n”. The class “c3” has instances “a”, “b”, and “c”, and the class “c4” has instances “g” and “h”. To do.

  In the ontology OA, the interest degree of the class “c3” is 1, the interest degree of the instance “a” is 2, the interest degree of the instance “b” is 5, and the interest degree of the instance “c” is 4. If there is, the degree of interest under the class “c3” is 1 + 2 + 5 + 4 = 12.

  Also, in the ontology OA, if the interest level of the class “c4” is 0, the interest level of the instance “g” is 2, and the interest level of the instance “h” is 1, the subordinate of the class “c4” The degree of interest is 0 + 2 + 1 = 3.

  In the ontology OA, if the interest level of the class “b2” is 1, and the interest degree of the instance “n” is 2, the interest degree under the class “b2” is 12 + 3 + 1 + 2 = 18.

  Also, in the ontology OA, if the interest level of the class “b1” is 0, the interest degree of the instance “m” is 3, and the interest degree of the instance “k” is 2, the subordinate of the class “b1” The degree of interest is 0 + 3 + 2 = 5.

  In the ontology OA, if the interest level of the class “a1” is 0, the interest level under the class “a1” is 5 + 18 = 23. In addition, classes “b1”, “b2”, and “b3” exist immediately below the class “a1” of the user B's interest ontology OB, and a class “c1” exists immediately below the class “b1”. And a class “c3” and a class “c4” exist immediately under the class “b2”, and an instance “l” exists in the class “c1”. The class “b2” has an instance “n”, the class “b3” has an instance “d”, the class “c3” has “a”, “c”, and It is assumed that an instance “e” exists, and a class “c4” includes instances “p” and “i”.

  In the ontology OB, the interest degree of the class “c3” is 0, the interest degree of the instance “a” is 3, the interest degree of the instance “c” is 3, and the interest degree of the instance “e” is 3. If there is, the degree of interest under the class “c3” is 3 + 3 + 3 = 9.

  Also, in the ontology OB, if the interest level of the class “c4” is 1, the interest degree of the instance “p” is 3, and the interest degree of the instance “i” is 2, the subordinate of the class “c4” Is 1 + 3 + 2 = 6.

  In the ontology OB, if the interest level of the class “b3” is 0 and the interest degree of the instance “d” is 2, the interest degree under the class “b3” is 0 + 2 = 2.

  Also, in the ontology OB, if the interest level of the class “b2” is 0 and the interest degree of the instance “n” is 6, the subordinate interest level of the class “b2” is 6 + 9 + 6 = 21.

  In the ontology OB, if the degree of interest of the class “b1” is 0, the degree of interest under the class “b1” is 0 + 3 = 3.

  In the ontology OB, if the interest level of the class “a1” is 0, the interest level under the class “a1” is 3 + 21 + 2 = 26.

Here, the terms used for calculating the degree of interest matching between ontology are defined as follows. Let T ₁ be the topology consisting of the parent class and child class, and T ₂ be the topology consisting of the class and instance. Furthermore, a common class of interest ontology OA and OB is defined as C _i, and a common instance is defined as I _i ∈I. In particular, a class set belonging to topology T ₁ is C (T ₁ ), and a class set belonging to topology T ₂ is C (T ₂ ). Further, the interest coincidence for the common class C _i is I (C _i ), the interest coincidence for the common instance I _i is I (I _i ), and the interest coincidence under the common class C _i is I _t (C _i). ).

(3-1) First, the common class C _i of the interest ontology OA and OB is analyzed, and the topology T ₁ is formed while paying attention to the presence of the common class C _i that forms both topologies T ₁ and T _2. extracts the common classes C _i to form a common class C _i and topology T _2. For example, in the example of FIG. 5, the common class a1, b1, b2 forms a topology T _1, the common class b2, c3, c4 can form a topology T _2.

(3-2) Assuming that the common class C _i has a common instance I _i between the interest ontologies OA and OB, the interest coincidence I (I _i ) with respect to the common instance I _{i is obtained} for the instances of the interest ontologies OA and OB. The lesser of interest. For example, in the example of FIG. 6, the degree of interest coincidence with respect to the common instance a under the common class c3 is 2 on the interest ontology OA side.

(3-3) The interest coincidence I (C _i ) for the common class C _i in the interest ontologies OA and OB can be the smaller of the interest degrees for the classes of the interest ontologies OA and OB. For example, in the example of FIG. 6, the interest coincidence with respect to the common class c3 is 0 on the interest ontology OB side.

(3-4) The interest coincidence I _t (C _i ) under the common class C _i ∈ C (T ₁ ) is the product set of the child classes under the class C _i in the interest ontologies OA and OB as N (C _i ), where the union is U (C _i ), it can be given by the following equation (4).

For example, in the example of FIG. 6, the interest coincidence I _t (C _i ) under the common class b2 is (9 + 3) / 2 = 6.

Then, as shown in the following equation (5), all common classes C _i value obtained by adding an interest matching score I _t (C _i) for forming a topology T _1, to form a topology T ₁ common The interest coincidence S (T ₁ ) for the class set can be used.

(3-5) On the other hand, the interest coincidence I _t (C _i ) under the common class C _i ∈ C (T ₂ ) is the instance set under the class C _i in the interest ontology OA as I _A (C _i ) If the instance set under the class C _i in the interest ontology OB is I _B (C _i ), it can be given by the following equation (6).

For example, in the example of FIG. 6, the interest coincidence I _t (C _i ) under the common class c3 is (2 + 0 + 3 + 0) / 4 = 5/4.

Then, as shown by the following equation (7), all common classes C _i value obtained by adding an interest matching score I _t (C _i) for forming a topology T _2, to form a topology T ₂ Common The interest coincidence S (T ₂ ) for the class set can be used.

(3-6) Then, assuming that the evaluation function corresponding to the importance with respect to the topologies T ₁ and T ₂ is f (X), the degree of approximation between the interest ontology OA and OB considering the degree of interest is
S _O (AB) = S (T ₁ ) + f (S (T ₂ )) (8)
Can be given in

  In this way, not only the extracted personal ontology topology but also the degree of interest of the user is taken into account, and the degree of approximation between the personal ontologies is measured, thereby taking into account the strength of the individual's interest and using only a character string. Instead, it is possible to search for information suitable for one's preference with high accuracy based on consistency between concepts.

  FIG. 7 is a diagram illustrating a community formation method using the approximation of interest ontology according to an embodiment of the present invention.

  In FIG. 7, it is assumed that the blog entries PA and PB of the users A and B are classified with respect to the template ontology, thereby generating the interest ontologies KA and KB of the users A and B, respectively (step S11). . Then, the degree of approximation between the interest ontologies KA and KB of the users A and B is measured (step S12), and the topology and the instance are analyzed by analyzing the classes and instances that co-occur between the interest ontologies KA and KB having a high degree of approximation. Although it is different, information that is highly likely to be of interest can be recommended to the user as unexpected information through the entry of another user (step S13).

  For example, by measuring the degree of approximation between the interest ontologies KA and KB of each user A and B, a user who is interested in a class such as “Madchester” or an instance such as “Happy Mondays” can have a class “Glasgow” It can be seen that there is a high possibility of being interested in an instance of “Teenage Fanclub”.

  Moreover, by applying such interest ontologies KA and KB to a blog, it is possible to support the formation of a community based on an unexpected entry recommendation based on the degree of approximation between the interest ontologies KA and KB rather than a simple keyword search. The user's interest can be naturally expanded (step S14).

<User Interest Correlation Graph Generator 20 (Step S200)>
(4) Next, based on the measured user interest similarity, the user interest correlation graph generator 20 draws an edge between users having similar interests with the user as a node, thereby creating a user graph (user interest correlation graph). Is generated. The graph generation process is executed according to the flowchart of FIG.

(4-1) In FIG. 8, first, in step S201, the similarity for each user measured by the similarity measurement method described in FIGS. Next, in step S202, a link is drawn with X users u _j having a high degree of similarity when viewed from the user u _i . The determination of this parameter X is important. After calculating a user who is interested in user A by the similarity measurement method described above, the types of instances possessed by the user are counted in the order of the user having the highest similarity. And, when the number of unique instances is added to a set of users with high similarity, if the number of instance types suddenly increases at a certain point, it is considered that the interest in user A has greatly shifted, The user up to that point is defined as a user set (consisting of X persons) with similar interests.

(4-2) Next, in step S203, a value obtained by normalizing the similarity of u _j viewed from the user u _i is given to the edge as a weight. This method of normalization is viewed from the user u _i, and the highest degree of similarity is low user similarity with a, when the highest similarity is larger user similarity is b, the edge between u _j and u _i The weight is
{Score (u _j ) −a} / (b−a) (9)
Given in.

<Recommended Score Calculator 30 (Step S300)>
(5) Next, the recommendation score calculator 30 causes the user's interest generated by the active user a to randomly walk on the user graph generated by the user interest correlation graph generator 20 starting from the item of interest possessed by the active user a. The recommended target list is calculated based on the number of appearances of the content and the edge score.

  The random walk uses graph-based collaborative filtering based on the random walk described in Chapter 3 of Non-Patent Document 2, but in the present embodiment, the elements of the Markov transition matrix in the random walk It implements by substituting the similarity of a user interest based on the similarity calculation based on the taxonomy described in FIGS.

<Recommended Item Affiliation Cluster Identification Device 40 (Step S400)>
(6) Next, the recommended item affiliation cluster identification unit 40 clusters the user graph generated by the user interest correlation graph generator 20, and includes the recommended item and background information of the item based on the clustering result. Create a recommendation list. The process is executed according to the flowchart of FIG.

  (6-1) In FIG. 9, first, in step S401, the user graph generated by the user interest correlation graph generator 20 is acquired. Next, in step S402, the method described in Non-Patent Document 3 is applied to the acquired user graph, and clustering is performed by, for example, the GN method.

  That is, for all edges in the user graph, betweenness (an index in which the inter-cluster edge has a high value and the in-cluster edge has a low value) is calculated, the edge having the highest betweenness is removed, and all the remaining edges Repeat the procedure for calculating betweenness again.

  Although the method described in the non-patent document does not consider the weight for the graph to which clustering is applied, the clustering in the present embodiment generates an appropriate cluster for the graph having a weight at the edge.

  Further, the new point in the clustering of the present embodiment is that the number of edges of the user graph can be controlled by the parameter X described in step S202 of FIG.

  (6-2) Next, in steps S403 to S406, a recommended item and background information of the item are generated using the result of clustering in step S402, and the recommended item and background information are generated by the recommended score calculator 30. The result is reflected in the determined recommendation target list and output as a recommendation list.

  That is, the cluster to which each user belongs is identified in step S403, the central user of each cluster is identified in step S404, the class possessed by each cluster is identified in step S405, and the recommended item and background information are added to the recommendation list in step 406. Reflect and output (present to user).

  The embodiment of the background information presentation of the recommended item is as follows.

  That is, using the clustering result, does the instance found at the random hop destination also exist in the cluster to which it belongs? If not, from which cluster is it discovered? Is specified.

  As an explicit method, first, (1) information indicating that it is found only from different clusters is presented. (2) As to which cluster it is from, the central user (or the instance set possessed by the user) having many edges in the discovery base cluster is exemplified to clearly indicate what information. Furthermore, since all the instance sets possessed by the cluster are presented with the amount of information received by the recommended user, the class information possessed by the cluster is presented.

  FIG. 10 shows an example of a recommendation list output from the recommended item belonging cluster identification device 40. In FIG. 10, “Artist a to d” are instances (recommended content) from the top to the bottom of the recommended ranking, and “class A does not have (or has)” in () indicates that the recommended content is It indicates whether or not the cluster to which it belongs also, “and cluster C2 to C4” indicates from which cluster it was discovered, and “central users B to E” are central users with many edges among the discovered clusters Represents a name.

  As in the embodiment of the present invention, the effect on the parity problem by using the user interest based on the taxonomy is as follows.

  It can be said that user interest based on taxonomy, particularly in the recommendation system, is a state where the system does not sufficiently grasp the user's interest (sparity problem is likely to occur) It can be said that it can be effectively used in a state in which only a part can be grasped) using the graph of FIG.

  FIG. 11 divides user interests ascertained by a certain information recommendation system into 10 types of situations where the system grasps from 1/10 to 9/10, and in each of the remaining 9/10 to 1/10 The system calculates whether or not the user can recommend a suitable content to the recommended user. In general collaborative filtering (see Non-Patent Document 5) used in Amazon and the like, and the taxonomy of the present invention The comparison is made with the recommendation based on the user interest and the similarity degree.

FIG. 11A shows comparison verifications when all users (3500) are targeted, and FIG. 11B shows comparison verifications when less interested users (2900) are targeted, and the solid line indicates a recommendation according to the present invention. FIG. 11A shows a recommendation error (when the X-axis is small), in particular, a recommendation error (when the X-axis is small). It can be seen that the average absolute error MAE) is small.

  Further, according to FIG. 11B, it can be seen that the MAE is always good when focusing on a user who has a small amount of user interest (the system cannot be grasped).

  From this, it can be seen that the method of the present invention is effective when the system does not sufficiently grasp the user's interest.

  Next, the recommendation accuracy of the recommendation system using the random walk in the user graph was verified using the user interest using the taxonomy of the present invention. Similarly to the above, the user's interest that a certain system grasps is divided into 10 types of situations that the system grasps from 1/10 to 9/10, and in each of the remaining 9/10 to 1/10, the system Calculated whether it was possible to recommend suitable content to the recommended users.

  In the conventional method using the instance-only matching, particularly when the system does not grasp the user's interest, the number of edges connecting the users is reduced. On the other hand, when a taxonomy is used, an edge may be drawn if the user is connected not only by an instance but also by a class.

  As a result of the verification, even in the graph base, in particular, when the restart parameter is set to about 0.4 to 0.6 (that is, restarts moderately) and the user's interest is 1/10 to 4 When only about / 10 was grasped, it was found that the method of the present invention had better accuracy. Furthermore, the accuracy (based on MAE) is better than in the case of the graph base. From this, it can be seen that the method of the present invention is suitable for discovering a user's potential interest in the graph base.

  Also, the present invention can be realized by configuring some or all of the functions of each means in the interest information recommendation apparatus of the present embodiment with a computer program and executing the program using the computer. It is needless to say that the procedure in the interest information recommendation method can be configured by a computer program and the program can be executed by the computer, and the program for realizing the function by the computer can be read by the computer, For example, FD (Floppy (registered trademark) Disk), MO (Magneto-Optical disk), ROM (Read Only Memory), memory card, CD (Compact Disk) -ROM, DVD (Digital Versati) e Disk) -ROM, CD-R, CD-RW, HDD, and recorded in a removable disk, or stored, it is possible or distribute. It is also possible to provide the above program through a network such as the Internet or electronic mail.

  The information communication system, which is one of the foundations of future industrial development, is currently undergoing rapid growth, but whether or not the information source that can be accessed by the information communication that has been expanded to an enormous amount can be fully utilized. It has become a big problem. The present invention personalizes personal information after easily creating a personal ontology and personalizing information that matches the interests of the individual, taking into account the interest of each individual and the strength of the individual's feelings about the subject. It is possible to recommend information and create a community based on the proximity of interest interests that have been created, and to automatically or efficiently utilize information that matches your interests from information sources of the information communication system There is sex.

DESCRIPTION OF SYMBOLS 1 ... Communication network 2-4 ... Terminal 5 ... Server 6 ... Model ontology 7-9 ... Blog site 10 ... User interest similarity measuring device 20 ... User interest correlation graph generator 30 ... Recommended score calculator 40 ... Affiliation of recommended item Cluster identifier

Claims (9)

A personal ontology that conceptually stratifies the interest information constructed for each user is generated, and the taxonomy is created in consideration of the class that includes the words in the interest information, the degree of interest for each instance, and the topology that is the connection form of the class. User interest similarity measurement means for measuring the interest similarity for each user based on;
Based on the user interest similarity for each user measured by the user interest similarity measurement means, a user interest correlation graph is generated by setting an edge between the users as a node and a weight based on the similarity as a score. A user interest correlation graph generating means,
Random walk on the graph generated by the user interest correlation graph generation means, add the content of interest of the user who has followed to the recommendation target list, based on the count number of each content and the score of the edge between the users A recommendation score calculating means for calculating a recommendation score of the content of interest;
Clustering the graph generated by the user interest correlation graph generating means, and based on the clustering result, a recommended item belonging cluster specifying means for creating a recommended list including recommended items and background information of the items;
An interest information recommendation device comprising: The user interest similarity measuring means includes:
A topology T ₁ composed of a parent class and a child class of each generated user ontology and a topology T ₂ composed of classes and instances are defined, and a common class is analyzed among a plurality of users' personal ontology to analyze the topology T _1. And the common class forming topology T ₂ are extracted,
Based on the product set and union of the child classes under the common class of the extracted two users' personal ontology, the interest coincidence under the common class is calculated, and the interest coincidence in all common classes forming the topology T ₁ The sum of degrees is calculated, and the interest matching degree S (T ₁ ) for the common class set forming the topology T ₁ is calculated.
Of the extracted common classes of two users' personal ontology, a common class is subordinate based on an instance set under a common class in one user's personal ontology and an instance set under a common class in the other user's personal ontology. The interest coincidence S is calculated, the sum of the interest coincidence in all the common classes forming the topology T ₂ is obtained, and the interest coincidence S (T ₂ ) for the common class set forming the topology T ₂ is calculated.
The degree of approximation So ontology between the ontology is calculated based on the calculated interest coincidence S (T ₁ ) and S (T ₂ ) and the evaluation function f (X) corresponding to the importance for the topologies T ₁ and T ₂ . The interest information recommendation device according to claim 1, wherein (T ₁ ) + f (S (T ₂ )) is obtained. The user interest correlation graph generation means includes:
Count the types of instances that each user has in order of the users with the highest interest similarity obtained by the user interest similarity measurement means, determine the counted number of users X immediately before the number of types of instances rapidly increases,
An edge is set between the X users uj having a high degree of similarity when viewed from a certain user ui,
The interest information recommendation device according to claim 1 or 2, wherein a value obtained by normalizing the similarity of the user uj viewed from the user ui is given to the edge as a weight. The background information of the recommended item in the recommended list created by the recommended item belonging cluster specifying means is whether the user's content of interest added to the recommended target list also exists in the cluster to which the recommended item belongs. 4. The information according to claim 1, including information indicating which cluster has been found when it does not exist, and information on a central user having many edges in the discovery base cluster. 5. The interest information recommendation device described. The user interest similarity measurer generates a personal ontology in which the interest information constructed for each user is conceptually hierarchized, and includes a class including words in the interest information, an interest degree for each instance, and a connection form of the classes. Considering the topology, a user interest similarity measurement step for measuring the interest similarity for each user based on the taxonomy,
The user interest correlation graph generation means sets an edge between users based on the user interest similarity for each user measured by the user interest similarity measurement means and having a user as a node and a weight based on the similarity as a score. A user interest correlation graph generating step for generating a user interest correlation graph;
The recommendation score calculation means randomly walks on the graph generated by the user interest correlation graph generation means, adds the content of interest of the user who has been traced to the recommendation target list, the count number of each content, and the edge between the users A recommendation score calculating step of calculating a recommendation score of the content of interest based on the score of
The recommended item affiliation cluster specifying unit clusters the graph generated by the user interest correlation graph generation unit, and creates a recommendation list including a recommended item and background information of the item based on the clustering result. Belonging cluster identification step,
An interest information recommendation method characterized by comprising: The user interest similarity measurement step includes:
A topology T ₁ composed of a parent class and a child class of each generated user ontology and a topology T ₂ composed of classes and instances are defined, and a common class is analyzed among a plurality of users' personal ontology to analyze the topology T _1. Extracting a common class forming a topology and a common class forming a topology T ₂ ;
Based on the product set and union of the child classes under the common class of the extracted two users' personal ontology, the interest coincidence under the common class is calculated, and the interest coincidence in all common classes forming the topology T ₁ Calculating an interest matching degree S (T ₁ ) for the common class set forming the topology T ₁ by obtaining a sum of degrees;
Of the extracted common classes of two users' personal ontology, a common class is subordinate based on an instance set under a common class in one user's personal ontology and an instance set under a common class in the other user's personal ontology. the step of interest matching degree calculated by the total sum of the interest matching score at all common classes of forming the topology T _2, to calculate an interest coincidence degree S (T ₂₎ to a common class set forming a topology T ₂ of the When,
The degree of approximation So ontology between the ontology is calculated based on the calculated interest coincidence S (T ₁ ) and S (T ₂ ) and the evaluation function f (X) corresponding to the importance for the topologies T ₁ and T ₂ . The interest information recommendation method according to claim 5, further comprising a step of obtaining (T ₁ ) + f (S (T ₂ )). The user interest correlation graph generation step includes:
Steps of counting the types of instances possessed by each user in descending order of the users' interest similarity obtained by the user interest similarity measurement means, and determining the counted number of users X immediately before the number of types of instances rapidly increases When,
Setting an edge with the X users uj having a high degree of similarity as viewed from a certain user ui;
Giving the edge a value obtained by normalizing the similarity of the user uj viewed from the user ui as a weight;
The interest information recommendation method according to claim 5 or 6, characterized by comprising: The background information of the recommended item in the recommendation list created by the recommended item belonging cluster specifying step is whether or not the content of interest of the user added to the recommendation target list exists in the cluster to which the recommended item belongs. 8. The information according to any one of claims 5 to 7, including information indicating which cluster is found when it does not exist, and information on a central user having many edges in the discovery base cluster. The interest information recommendation method described. An interest information recommendation program for causing a computer to function as each means according to any one of claims 1 to 4.

Images